LingBot-Map runs out-of-the-box on AMD MI300X with ROCm 6.4 (`--use_sdpa`)

Hi, lingbot-map team, thanks for this great work, just verified on AMD MI GPUs and it run through well. thanks for adding AMD GPU in supported list if possible.

Summary

LingBot-Map works out-of-the-box on AMD Instinct MI300X GPUs with ROCm 6.4, using the existing --use_sdpa fallback path. No code changes are required.

This issue documents the verified setup and results for the community.

Test Environment

Installation on ROCm

# 1. Start from ROCm PyTorch Docker image
docker run --rm -it \
  --device=/dev/kfd --device=/dev/dri \
  --ipc=host --security-opt seccomp=unconfined \
  rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0 bash

# 2. Clone and install (identical to CUDA setup)
git clone https://github.com/robbyant/lingbot-map.git
cd lingbot-map
pip install -e ".[vis]"

# 3. Download checkpoint
hf download robbyant/lingbot-map lingbot-map.pt --local-dir checkpoints/

# 4. Run with --use_sdpa (FlashInfer is not available on ROCm)
python demo.py \
  --model_path checkpoints/lingbot-map.pt \
  --image_folder example/church \
  --use_sdpa

Results

Tested on the built-in example/church scene (286 frames):

What works

• pip install -e ".[vis]" — all dependencies install cleanly, no CUDA-specific blockers
• Model loading (1.19B params, 4.4 GB checkpoint)
• Streaming inference (--mode streaming) with --use_sdpa
• bfloat16 autocast
• Depth prediction + world point estimation + pose estimation
• Point cloud export

What doesn't work (expected)

• FlashInfer — no ROCm/HIP support currently.

some results

Start with the project’s existing supported-GPU documentation and compare it with the verified setup in this issue. Use demo.py with the example/church scene and --use_sdpa to confirm the documented path on AMD MI300X with ROCm 6.4. Done means the support information includes the working environment, command, results, and FlashInfer limitation.